StrokeScope is a web application that uses a convolutional neural network (CNN) to analyze CT brain scans and detect the presence of hemorrhagic strokes. Users upload a CT scan image, and the application preprocesses it, runs it through a trained machine learning model, and returns a classification indicating whether a hemorrhage is present, a confidence score, and a highlight of the problematic area. The app also recommends next steps based on the result, making it useful for individuals without a medical imaging background.
StrokeScope was built by Abhi S., Sahasra C., and Saara P.
Stroke is one of the leading causes of death and long-term disability globally. One in four people will experience a stroke above the age of 25. Every minute a stroke goes untreated, the brain can lose an estimated 1.9 million neurons — making early detection critical to survival and recovery.
Despite this urgency, long wait times, limited access to specialists, and financial barriers regularly prevent patients from receiving timely imaging analysis. Many healthcare centers, particularly in underserved areas, lack access to dedicated imaging software or trained radiologists available around the clock. The result is delayed diagnosis and preventable damage. StrokeScope was built to lower the barrier to early scan analysis for anyone with access to a browser.
StrokeScope is designed for three primary groups:
General public: Individuals or family members who have obtained a CT scan but lack the medical knowledge or time to interpret it, and want a fast, accessible first read before consulting a physician.
Students and researchers: Those looking to experiment with machine learning and CT scan analysis in a hands-on environment, using real imaging data and a working model pipeline.
Healthcare centers: Clinics and hospitals without access to dedicated imaging software or immediate radiology support that need a lightweight, cost-free screening tool to assist triage.
StrokeScope is a full-stack web application with three main components: a React/Dart frontend that renders the user interface, a Python backend that handles model inference and image processing, and an API layer built with FastAPI that connects the two. The CNN model was trained on 40% of the RSNA dataset for 18 epochs and achieves an average AUC score of 0.0087 across hemorrhage classes. Saved model files are transferred to the API layer, which classifies uploaded scans and returns structured results to the frontend for display.
The hemorrhagic stroke classification task is particularly challenging because hemorrhages vary significantly in size, location, and subtype. StrokeScope addresses this by training on multiple hemorrhage classes simultaneously, allowing the model to identify a range of presentations rather than a single injury pattern.
The MVP of StrokeScope delivers three core features working end to end:
Stroke Detection
The application uses a convolutional neural network trained on publicly available medical imaging datasets to analyze a CT brain scan and classify whether a hemorrhage is present or not. The model outputs a classification result along with a per-class confidence score.
Image Processing
Users upload CT brain scan images directly through the application interface. The system automatically handles image normalization and preprocessing prior to analysis, ensuring inputs are correctly formatted for the model regardless of the original scan format or resolution.
Displaying Results
The app indicates whether a hemorrhage is present, displays the model's confidence level for each hemorrhage class, and highlights the problematic areas within the scan. Based on the result, the app recommends the appropriate next steps for the user to take.
The poster below was presented at the Apps for Good showcase. You can scroll through it or open it in full screen.